VP9

VP9 is a royalty-free video format developed by Google as the successor to VP8 and released in June 2013 as part of the WebM Project. It competes with HEVC: efficiency is in the same range, decoding is relatively light, and licensing is simpler. YouTube adopted it early and still serves much of its catalogue in VP9; it is also widely used in WebRTC, where its scalable (SVC) mode suits multi-party video calls.

VP9 was never put through a standards body. The bitstream has been frozen since 2013, and the WebM Project’s published specification is still labelled draft v0.7 (2017).

When to use VP9

• Web delivery to browsers and devices that lack AV1 decode. Nearly every browser plays VP9, including Safari (macOS since Safari 14.1, iOS since 17.4).
• HDR on the web or YouTube via profile 2 (10-bit), where HEVC is not an option.
• For new encodes where AV1 decode is available, AV1 is more efficient; VP9 remains a solid middle rung between AV1 and H.264.

Profiles and levels

Profile	Bit depth	Chroma
0	8	4:2:0
1	8	4:2:2, 4:4:0, 4:4:4
2	10, 12	4:2:0
3	10, 12	4:2:2, 4:4:0, 4:4:4

RGB is coded as 4:4:4 in profile 1 or 3. Profile 0 is what every decoder supports; profile 2 at 10-bit is the usual HDR profile and is widely hardware-decoded on recent devices. Profiles 1 and 3 are mostly software-only.

Levels run from 1 to 6.2 and limit sample rate, picture size, bitrate and tile columns. They matter mostly for hardware decoders.

Coding tools

VP9 sits between VP8 and AV1 in design:

• 64×64 superblocks, recursively partitioned down to 4×4.
• 10 intra modes: DC, vertical, horizontal, six diagonal directions and TrueMotion.
• Three active references per frame (last, golden, alt-ref), chosen from eight stored frames, with compound prediction and 1/8-pel motion vectors using regular, sharp or smooth interpolation filters.
• Alt-ref frames: hidden frames built from future (often temporally filtered) source frames and used only as references. They do much of the job of B-frames without frame reordering; libvpx’s --auto-alt-ref and --lag-in-frames control them.
• DCT and ADST transforms (mixed per direction) from 4×4 to 32×32, and a lossless mode using a Walsh-Hadamard transform.
• A boolean arithmetic coder whose probabilities adapt between frames.
• Segmentation (up to 8 segments with their own quantiser, loop-filter level, reference or skip), a deblocking loop filter, tiles for parallel encoding and decoding, and reference scaling so resolution can change without a keyframe.

Encoders and decoders

VP9 implementations

Relationship	Project or format	Evidence and scope	Status
Implements by	libvpx / vpxenc v1.17.0	libvpx v1.17.0 provides VP9 encoding and decoding.Scope: Encoder Decoder · Sources: libvpx v1.17.0 source, VP9 bitstream and decoding process specification	Documented
Implements by	SVT-VP9 v0.3.1	SVT-VP9 v0.3.1 provides an experimental VP9 encoder.Scope: Experimental Encoder · Sources: SVT-VP9 v0.3.1 repository	Documented

• libvpx (vpxenc, or libvpx-vp9 in FFmpeg) is the reference and practically the only software encoder in use. High-bit-depth (profile 2/3) encoding needs a build with high-bit-depth support enabled; most FFmpeg builds have it.
• SVT-VP9 and Eve-VP9 are alternatives with limited adoption.
• Hardware encoding: on PCs only Intel GPUs encode VP9 (Quick Sync, VA-API); NVIDIA NVENC and AMD AMF do not. Some Android SoCs (Qualcomm, Samsung Exynos) expose VP9 encoders through MediaCodec, mainly used by WebRTC.

A constant-quality VP9 WebM with FFmpeg:

ffmpeg -i input.mkv -c:v libvpx-vp9 -crf 31 -row-mt 1 -c:a libopus output.webm

With -crf and no bitrate, FFmpeg runs libvpx in constant-quality mode (older guides add -b:v 0 for the same effect; with a non-zero -b:v, -crf becomes constrained quality with that bitrate as a cap). -row-mt 1 enables row-based multithreading, which libvpx leaves off by default and which speeds up encoding considerably on multi-core machines. Lower CRF means higher quality; see vpxenc for speed (-cpu-used), two-pass and tile settings.

Decoders: FFmpeg’s native vp9 decoder is fast and is what most players use; libvpx also decodes. Hardware decode of profile 0 is near-universal on current GPUs and phones; on NVIDIA it starts with the GTX 950/960 (8-bit only) and 10/12-bit with Pascal (GTX 10 series). See video codec support for browsers.

Containers and HDR

VP9 container mappings

Relationship	Project or format	Evidence and scope	Status
Carried By	WebM	WebM's documented codec subset includes VP9.Scope: Container Subset · Sources: WebM container guidelines	Documented
Carried By	ISO Base Media File Format	The WebM Project publishes a VP codec binding for ISO Base Media File Format.Scope: Container Mapping · Sources: VP Codec ISO Media File Format Binding	Documented
Carried By	Matroska	The Matroska mapping uses the V_VP9 Codec ID.Scope: Container Mapping · Sources: Matroska codec mappings	Documented

• WebM / Matroska (V_VP9) is the normal container.
• MP4 via the WebM Project’s VP codec ISOBMFF binding (vp09 sample entry). Browser MSE and DASH players support it, but some players and editors expect VP9 only in WebM.
• IVF is the simple raw format vpxenc writes for testing.

VP9’s own header carries bit depth, chroma subsampling, range and a colour-space value. It has no field for the transfer function (PQ or HLG), so HDR VP9 relies on the container’s colour metadata: WebM/Matroska Colour elements, or in MP4 the vpcC box (primaries, transfer, matrix) plus SmDm/CoLL boxes for mastering metadata. Check that your muxer writes them. HDR10+ metadata can be carried in WebM as ITU-T T.35 data.

Licensing

Google released VP9 royalty-free with an additional patent grant for its own patents, and libvpx is BSD-licensed. Sisvel has offered a VP9 patent pool to device makers since 2019, claiming patents from third parties not bound by Google’s grant.

Pitfalls

• Slow encoding at default settings: libvpx’s best-quality modes are slow; use -row-mt 1 and a higher -cpu-used for faster turnaround.
• Profile 1/3 output: 4:4:4 or RGB input produces profile 1 or 3 streams that many hardware decoders reject. Use -pix_fmt yuv420p (or yuv420p10le for HDR) for delivery.
• HDR without metadata: a 10-bit profile 2 stream with no transfer and primaries in the container plays as SDR with wrong colours.